Single Demo
You are hereOne question at a time. See each agent step-by-step.
Batch Demo
Training ModeProcess many samples. Watch the Playbook learn!
What is ACE?
Agentic Context Engineering
ACE is a self-improving AI framework where LLMs learn from their mistakes without changing their weights. Instead of retraining, ACE builds an evolving playbook—a living document of strategies, insights, and lessons learned that grows smarter with each problem solved.
Context adaptation instead of weight updates—preserves knowledge without catastrophic forgetting.
+10.6% on agent tasks, +8.6% on finance—matches production agents with smaller models.
Learns from execution feedback—no labeled data required for continuous improvement.
The Three-Role Agentic Architecture
Generator
Step 1
Reads the Playbook and applies its strategies to answer the question. Reports which bullet points it used in its reasoning.
Reflector
Step 2
Analyzes the Generator's output. Tags each bullet as helpful, harmful, or neutral. Identifies root causes of errors.
Curator
Step 3
Updates the Playbook with new insights from the Reflector's analysis. Uses incremental delta updates to prevent context collapse.
Powered by SCX.ai & DeepSeek-V3.1What is the Playbook?Key Concept
The Playbook is ACE's secret weapon — it's not just a prompt. It's a living knowledge base that grows and improves with every problem the system solves.
- • Static — never changes
- • Written by developers
- • Same for every task
- • Dynamic — evolves over time
- • Updated by the Curator agent
- • Tracks what works (helpful/harmful counts)
Example Playbook Entry
[str-00001] helpful=3 harmful=0 :: Look for capitalized words to identify proper nounsstr-00001
+3 helpful
The actual insight
💡 Think of it like: A prompt is instructions for a new employee. The Playbook is the knowledge base they build up over months of experience.
Behind the Scenes: How ACE Calls the LLMTechnical
The three "agents" (Generator, Reflector, Curator) are not separate servers. They're just 3 different prompts sent to the same LLM.
Each Problem = 3 LLM API Calls
"Answer this question"
"What went wrong?"
"Update playbook"
Same LLM API(DeepSeek-V3.1 via SCX.ai)3-7 LLM calls per problem. Expensive, but builds up the Playbook with knowledge.
1 LLM call per problem. Just the Generator with a pre-built Playbook. Fast & cheap!
Where Does Everything Run?
On YOUR Server:
- • Generator, Reflector, Curator code
- • Playbook storage (database/file)
- • Orchestration logic
On LLM Provider (SCX.ai):
- • The actual AI model
- • Stateless — no memory between calls
- • Just processes prompts
⚡ Key insight: The "agents" are just different prompts. All the intelligence comes from how you orchestrate the calls and persist the Playbook.
Select a Problem
Current Playbook
What is a Playbook?
The Playbook is ACE's "living memory." It contains strategies and lessons learned from past tasks.The Generator uses these bullets to avoid mistakes and follow best practices. The Reflector evaluates if they helped or hurt. The Curator adds new insights to this list in real-time.
Living Knowledge: This list is passed to the AI as context. Watch it grow! The Curator agent will add new rules here based on what it learns from the Reflector's feedback.
ACE Agent Pipeline
PaperGenerator
Waiting...
Reads the Playbook and uses its strategies to answer the question. Reports which bullet points it referenced.
Reflector
Waiting...
Analyzes the Generator's reasoning. Tags bullets as helpful/harmful. Identifies the root cause of any errors.
Curator
Waiting...
Updates the Playbook with new insights. Uses incremental delta updates to prevent context collapse.